A Deep Reinforcement Learning Framework for Automatic Operation Control of Power System Considering Extreme Weather Events

With the continuous integration of intermittent renewable energy and large-scale regional interconnection of power systems, the power system has evolved into a high dimensional complex nonlinear system. Undoubtedly, in such a highly complex system, more intelligence and flexibility for the operation, control and decision-making would be required. Therefore, a novel deep reinforcement learning (DRL) framework for automatic operation control (AOC) of the power system considering the existence of extreme weather events is proposed in this paper. In order to reduce the destructive impact of extreme weather events, topology switching control must be utilized except for the generator redispatching and load shedding, which poses a huge challenge to the optimization of the DRL algorithm. To tackle this problem, a novel action space reduction method which utilizes the domain knowledge, historical data and heuristic constraints, is firstly proposed. Then, imitation learning (IL) is introduced to pre-train the DRL agent and build a feasible topology database for extreme weather events. Finally, the Soft Actor-Critic (SAC) algorithm with improved exploration and policy update strategy is developed to train the agent. Numerical studies in a modified 36-bus system indicate that the proposed model and method have good convergence characteristics and solving efficiency.